-
Freelance agents introduce new financial risks, according to the company.
-
The report notes that current libraries were not designed for AI without human supervision.
The advancement of artificial intelligence (AI) is transforming software development. A report from the Argentine company Lambda Class points out that this change generates opportunities and risks in the cryptocurrency ecosystem, especially when automated systems interact directly with real money without constant human intervention.
In the document published on January 23, the company focused on the development of tools for Ethereum proposes that the use of AI agents to operate with cryptocurrencies introduces new vectors of security flaws. These are elements that were not contemplated in the original design of the infrastructure.
According to the report, the introduction of AI agents (programs capable of making decisions and executing actions autonomously) alters an important premise that is part of the design of Ethereum. This is because its general purpose financial infrastructure is based on operations are initiated and understood by human persons.
Therefore, when AI systems interact directly with the network and sign transactions without prior human review, errors no longer remain at the conceptual level, but rather translate into immediate and irreversible economic losses.
The Lambda Class team’s analysis takes on special relevance given that on January 29 The ERC-8004 standard was implemented on the Ethereum main network. As reported by NoticiasVE, this standard would precisely provide Ethereum with a system in which AI agents can connect, verify and repute each other automatically through smart contracts.
What if AI replaces the human operator?
According to the Lambda Class report, libraries (software toolkits that developers use to interact with Ethereum and send transactions) were designed for people, not for autonomous systems.
Tools like ethers.js or web3.js assume that someone understands what they are signing before authorizing a transaction. That model, as stated above, could fail when the operator is an AI:
- an agent can hallucinate an addressthat is, generating a valid but incorrect address.
- Can confuse units. For example, interpreting “send 100, as 100 ethers instead of 100 dollars”
- You can also be manipulated through instruction injection, a technique that introduces malicious commands into the data it processes.
Each of these errors is unlikely in isolation. However, the report warns that when millions of automated trades are executed, these failures they become inevitable.
In Ethereum there is no bank that reverses operations. Once a transaction is confirmed, funds are permanently lost (except in the famous The DAO hack).
Lambda Class emphasizes that this is not a “improving AI” problem. The risk arises from allowing imperfect systems operate directly on irreversible financial infrastructure. When something fails, the system returns technical messages that an AI cannot safely interpret.
The report compares this scenario to letting a robot drive a truck without automatic brakes: The problem is not the intention of the agent, but the absence of barriers that stop him when something goes wrong.
Restrictions as a layer of defense
To address this problem, the Lambda Class team believes that the way to reduce risks is not by making AI “smarter”, but by put structural limits.
For that, he developed eth-agent, a development kit that introduces mandatory restrictions in the execution of transactions in each wallet. For example, spending caps per transaction, per hour and per day. In this way, if an agent tries to exceed these limits, the operation automatically failswith no possibility of evasion.
The system also returns clear and structured errors. Instead of difficult-to-interpret technical messages, it informs you which rule was violated and when it is safe to retry.
Additionally, for sensitive transactions (such as large amounts or new recipients) requires human approval before executing the shipment.
There are ways to avoid the risks of AI
As part of the forecasts, the study advises that self-employed agents operate mainly with stablecoinsin order to avoid errors caused by price volatility.
It also recommends incorporating smart accounts under the ERC-4337 standard, which allow delegate permissions in a limited and controlled way.
The central idea of these proposals is similar to that of an operating system. Applications may crash, but the core imposes rules that prevent further damage. In decentralized finance, that “core” must protect even when the AI makes mistakes.
The report concludes that AI agents will continue to improve, but they will never be perfect. In a financial system without error reversal, relying on their correction is insufficient.